A cover letter that matches the vacancy perfectly. A CV without weak spots. Answers in a video interview that sound as if they were thought through in advance. Increasingly, they were, with help from AI. For anyone who selects people, that turns an old question into an urgent one: which part of your selection process still measures something about the person?
What actually predicts well?
For decades, the overview by Schmidt and Hunter from 1998 was the reference point. Cognitive ability tests, work samples and structured interviews emerged as the strongest predictors, all around .50. In 2022, Sackett, Zhang, Berry and Lievens published a reanalysis in the Journal of Applied Psychology. Their conclusion: because a statistical correction was applied too generously, most predictors have been overestimated for years, on average by .10 to .20 points.
Their revised ranking, expressed as the relationship with job performance (where 1 is a perfect prediction):
- Structured interview: .42
- Job knowledge test: .40
- Work sample: .33
- Cognitive ability test: .31
- Assessment centre: .29
- Unstructured interview: .19
- Conscientiousness, measured with a general personality questionnaire: .19
- Years of work experience: .07
Three things stand out. The structured interview comes out on top. The free-form getting-to-know-you conversation predicts only half as well. And years of experience say almost nothing about how someone will perform. No instrument comes close to a perfect prediction, so relying on a single method is a big risk.
What AI changes
Those figures come from a time when candidates gave their own answers. That can no longer be taken for granted. In an experiment with 245 participants, people who used ChatGPT in an asynchronous video interview, where you answer questions at home in front of a camera, received considerably higher scores on content and overall performance than people without AI (Canagasuriam & Lukacik, 2025). Research on personality questionnaires shows that ChatGPT can steer a profile towards a desired role, although without extra instruction it did not consistently outperform people who tried to present a favourable picture themselves (Robie et al., 2025).
The lesson: components a candidate completes alone and unsupervised are becoming more vulnerable. They remain useful, for example for an initial screening of large numbers. You just can no longer let the weight of an appointment rest on them.
For practice, AI is remarkably strong: you can redo a difficult conversation ten times, with immediate feedback. Judging a person calls for something else: behaviour that was not scripted in advance, and someone who can weigh it.
What this means for your selection
1. Structure your interviews. Ask every candidate the same role-specific questions, probe for concrete behaviour and assess against criteria agreed in advance. That is the difference between .42 and .19.
2. Let people show what they can do. A simulation in which someone responds live to a difficult employee or a critical client leaves no room for a chatbot to feed the answers. You see behaviour instead of a prepared text.
3. Combine methods. No single instrument predicts strongly enough on its own. A structured interview, a cognitive ability test and a behavioural simulation complement each other and together give a sturdier picture.
4. Keep a psychologist in the assessment. Someone who has actually seen and spoken with the candidate notices when an answer is too polished, and then asks further.
How we approach this
In our selection assessments we combine COTAN-certified personality and cognitive ability tests with behavioural simulations and an in-depth interview. One experienced consultant guides the candidate through the whole day. We use online tests for pre-selection with larger groups, always combinable with an in-depth assessment on location. The psychologist remains in charge of the judgement.
Wondering whether your selection process still measures what you want it to measure? See our assessments, or read what the EU AI Act requires for human oversight in HR decisions.
Sources
- Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040-2068. doi.org/10.1037/apl0000994
- Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings. Psychological Bulletin, 124(2), 262-274. doi.org/10.1037/0033-2909.124.2.262
- Canagasuriam, D., & Lukacik, E.-R. (2025). ChatGPT, can you take my job interview? Examining artificial intelligence cheating in the asynchronous video interview. International Journal of Selection and Assessment, 33(1). doi.org/10.1111/ijsa.12491
- Robie, C., Phillips, J., Bourdage, J. S., Christiansen, N. D., Dunlop, P. D., Risavy, S. D., & Speer, A. B. (2025). Can ChatGPT outperform humans in faking a personality assessment while avoiding detection? International Journal of Selection and Assessment, 33(3). doi.org/10.1111/ijsa.70015